Kimi K3 API: how to get a key, what it costs per 1M tokens, and whether it's free

Kimi API pricing at a glance

Moonshot's own pricing page, read on 29 Oct 2026, in USD per 1M tokens. All K3-series models have a 1,048,576-token context window.

Three things on that page matter before you pick a model:

• K3 is no longer the newest. Moonshot's K3 guide calls it "Kimi's previous-generation flagship model". The current one is Kimi K3.1 Pro, which is cheaper per token. Moonshot's billing FAQ even says: "If cost is a concern, you can use the kimi-k3.1-pro model."
• Caching is on by default. "The Kimi API automatically enables context caching billing for repeated request prefixes, with a default TTL (time-to-live) of 5m." A long system prompt you resend is billed at the cached rate, a tenth of the input price.
• K2 is closing to new accounts. "Starting October 31, 2026, K2 series models will no longer be available to newly registered users", and Moonshot plans to retire all K2 models by the end of December 2026. Build on K3 or K3.1 Pro.

How to get a Kimi API key

Moonshot's quickstart puts it in one line: "Visit the Kimi API Platform, sign in, then go to API Keys to create and copy an API Key." Step by step:

1. Sign in at platform.kimi.ai. The platform supports two sign-in methods: a verification code sent to your email, or "Sign in with Google".
2. Top up at least $1. The K3 guide says: "Kimi K3 is a flagship model: it is unlocked after a successful top-up (minimum $1)." Moonshot's billing FAQ, read on 29 Oct 2026, lists "WeChat Pay and Alipay QR code payments" for individual online top-ups, so check the top-up page for the payment methods your account is offered before you plan around a card.
3. Create the key on the API Keys page and store it in an environment variable. Moonshot's own samples call it MOONSHOT_API_KEY.
4. Point your code at Moonshot. The API speaks the OpenAI format at https://api.moonshot.ai/v1 and the Anthropic Messages format at https://api.moonshot.ai/anthropic. The model name for K3 is kimi-k3.

A minimal call with the OpenAI Python SDK, from Moonshot's K3 guide:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

completion = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Introduce Kimi K3 in one sentence."}],
)
print(completion.choices[0].message.content)

K3 always thinks before it answers. You can set how hard with the top-level reasoning_effort field, which takes low, high or max, and the default is max. Since output tokens cost five times as much as input, low is worth testing for simple tasks.

Is the Kimi API free?

No. On 29 Oct 2026 we read Moonshot's pricing page, rate-limit page, quickstart, K3 guide and billing FAQ, and none of them offers a free tier or a trial. Every request is billed per token from a prepaid balance.

What the rate-limit page does say: "a minimum recharge of $1 is required to start using the API; when your cumulative recharge reaches $5, you receive a $5 voucher." So the closest thing to free credit is $5 back after your first $5. Vouchers "do not count towards the cumulative recharge total", which matters for the next point.

Your top-ups also set your rate limits. Moonshot ties limits to how much you have added in total, not to what you spend:

At Tier0 you get one request at a time and three a minute. That is fine for testing and too slow for a coding agent. If you plan to point a tool like Claude Code at Kimi, budget for at least the $10 tier. Moonshot publishes its own Claude Code setup guide, which uses https://api.moonshot.ai/anthropic as the base URL.

Searches for a "free Kimi API" turn up the same three things as for any popular model. The open weights are free to download, but nobody can run them at home (more below). Third-party "free" endpoints send your prompts to whoever runs them. And a "free Kimi API key" from a website belongs to someone else's account: it can be revoked at any time, and everything you send runs through an account you don't control. Don't use one, and never pay for one.

Where your data goes

Moonshot's privacy policy (last updated 30 Apr 2025) says the platform is run by Moonshot AI Pte. Ltd. in Singapore and states: "We store the information we collect in secure servers located in Singapore." It also says personal information is used for "training and refining our underlying technology, such as machine learning models and algorithms." Its Zero Data Retention page says Moonshot "does not use enterprise customer data to train models" by default, and offers ZDR, where prompts are deleted once the response is returned, "on request to enterprise customers."

If your prompts include client data or code you can't share, read those two pages before you start, or ask Moonshot's sales team about ZDR.

Kimi K3 on Novita AI: same price, one key for many models

Moonshot published K3's weights, so other companies can serve the same model. Novita AI is one of them, and it lists K3 under the model id moonshotai/kimi-k3. You sign up with Google, GitHub or email, create a key, and call an OpenAI-compatible endpoint at https://api.novita.ai/openai/v1. Novita also lists an Anthropic-compatible endpoint for its Kimi models.

Prices per 1M tokens, read on 29 Oct 2026 from Moonshot's pricing page and Novita's public model list:

Novita is not cheaper for Kimi. The per-token prices are identical. Use Moonshot directly if Kimi is the only model you need and its sign-up, payment and data terms work for you.

Novita makes sense for other reasons:

• One key for many models. Novita's public list has 99 entries on 29 Oct 2026, including DeepSeek, GLM, Qwen, Llama and gpt-oss next to Kimi. You can switch models by changing one string.
• Sign-up and billing. Novita takes Google, GitHub or email sign-up, and you don't need a Moonshot account or its top-up methods.
• Not a Moonshot account. Your prompts go to Novita instead of Moonshot. Novita's privacy policy is written for US privacy law but does not say where inference runs, so ask Novita before you send anything sensitive.

Disclosure: Novita AI links on this page are referral links. If you sign up through one, we earn a commission on what you spend there. Moonshot pays us nothing, and the prices above are the same on both.

Can you run Kimi K3 yourself instead?

Not at home. K3 has 2.8 trillion parameters, and its official MXFP4 weights total about 1.56 TB. The smallest community quant is still 509 GB, and vLLM's recipe asks for at least eight GB300 GPUs. Our Kimi K3 hardware breakdown has the full memory math. Moonshot says the weights for K3.1 Pro "will be released by Nov 4", and at 3T class it will not be any smaller.

So for one person or a small team, the API is the only realistic way to use K3. If you want a model that runs on your own card, with no key and no per-token bill, see the open models in what fits a 16GB GPU or the best free local AI models.

Keep your Kimi API key safe

A Kimi key spends your balance, and Moonshot enforces rate limits "at the user level, not the key level", so a leaked key also eats the limits your own app needs.

• Server side only. Moonshot's quickstart: "Keep your API Key secure. Do not share it with others or hard-code it directly in your application." If your app is a Lovable, Bolt or Next.js front end, the call has to go through a server function. Your API keys are in the browser shows how.
• Keep it out of git. Put it in an environment variable and keep .env out of your repo. Watch out for VITE_ and NEXT_PUBLIC_ prefixes, because they ship the value to the browser. The one file you must never commit explains why.
• Set a daily budget. Moonshot's billing FAQ recommends a project daily spending budget in the project settings: "Once the budget limit is reached, the system will automatically reject all API requests under that project." It warns the limit "may take about 10 minutes to take effect", so it caps damage rather than stopping it instantly.
• If it leaks, revoke it on the API Keys page, create a new key and swap it into your app.

Common questions